Meta's AI agent Muse quickly attracted attention after launch and reached number one in the Apple App Store. Different figures are circulating about user numbers: the headline of one report mentions 50,000 users in the first week, while the text of the same report mentions over 500,000. The order of magnitude therefore remains uncertain – all that can be said for certain is that the launch generated an extraordinary amount of interest.

More interesting than the number are two points that became public shortly after launch. They touch on questions that regularly arise in corporate projects involving AI: where does the code come from, and what does a so-called agent actually do itself?

Point one: origin of the code

Meta admits that Muse is "strongly inspired" by the open-source project OpenClaw. File names and contents are said to be partly almost identical. This is more than a matter of style. Open-source software is not the same as freely usable: every licence ties use to conditions, from naming the author to the obligation to publish derivative works under the same licence.

Anyone who adopts third-party code – whether manually or via an AI model trained on public repositories – potentially also adopts these conditions. For Meta, this is a reputational issue. For a mid-sized company that ships an application or passes it on to customers, it can become a concrete legal risk.

Point two: how autonomous is the agent?

According to its product promise, Muse is supposed to make calls on behalf of users, among other things. Internal communication suggests that these calls are instead handled by call-centre staff. The matter has not been conclusively confirmed; the reporting is based on internal documents.

The pattern is not new. Between "the system can do that" and "the system does that on its own in normal operation" there is often a piece of human work – as a fallback, for quality assurance, or because the last few per cent of reliability are technically expensive. It becomes problematic where this share is not communicated. Because users make their decisions – for example about which data they disclose – on the basis of the assumption that they are talking to a machine.

What follows from this for your own projects

Both points can be translated into concrete requirements that arise in every project in which AI tools or agent systems are used.

Clarify the origin of code

  • Licence review as a fixed step: Dependencies and adopted code fragments should be documented and checked for their licence – automated in the build process, not only at audit time.
  • Label AI-generated code: If assistant systems are in use, an internal rule helps to define which kinds of suggestions may be adopted and where manual review is needed.
  • Contractual assurances: For purchased software, it is worth asking what warranty the provider gives regarding the chain of rights.

Describe autonomy realistically

  • Separate demo from normal operation: What works in a controlled demonstration is not yet a robust process. A pilot phase with real cases and a measured error rate makes sense.
  • Name the human share: A human-in-the-loop – i.e. a person who reviews or approves intermediate results – is not a flaw, but often the only defensible way of building it. However, it belongs in the process documentation and, where third parties are affected, in communication.
  • Think about data protection: If people see content or hold conversations that users believe are processed by machine, this changes the data protection assessment.

Hype as a pace-setter

The fact that OpenAI is already discussing its own response to Muse shows the dynamics in the field. Such races create pressure to announce features early and to hide the gap between announcement and implementation behind the product name. For companies that want to buy or integrate such services, this is an argument for sobriety: the basis for planning is not the announcement, but what can be reproduced in your own test.

A few clear questions to providers help here: Which steps run fully automated? Where do humans intervene? Which data leaves the system in the process? And how is the chain of rights to the code documented? Answers that remain evasive are themselves information.

Assessment

For web and software projects, the Muse case means above all two things. First, checking licences and code origin belongs in the standard workflow as soon as AI tools are involved in creation – doing it retrospectively is considerably more costly. Second, agent functions in an offering should be described as they actually run in operation, including the points at which humans review or intervene. This protects against disappointment in the project and against discussions that nobody wants to have later.

Sources